NVIDIA Opens Vera Rubin Rack Designs at OCP: MGX 1.2, DSX Software and Grid Specs

At the OCP Global Summit on October 12, 2026, NVIDIA said it is opening its Vera Rubin rack architecture to the wider industry: the rack and compute tray designs are being contributed to the Open Compute Project (OCP) through the MGX modular architecture. According to NVIDIA's OCP announcement, the package also includes the MGX 1.2 specification, a Vera CPU rack spec, four open-source DSX software components, new Spectrum-6 networking, USFP optics, and two facility-level standards co-published with hyperscalers.

For anyone tracking where AI compute is going, this is less about a single chip and more about who gets to build the box around it. Open designs lower the cost for server makers, cooling vendors and data center operators to ship Vera Rubin-class racks, and that shapes how fast capacity appears.

TL;DR: what was opened, and why you might care

MGX 1.2: the rack gets bigger and wetter

NVIDIA describes MGX 1.2 as the third-generation MGX reference design delivered through OCP's Open DC Initiative. The headline numbers are physical: rack deployments up to 250kW, a liquid-cooled 5,000A busbar, rack-scale blind-mate liquid cooling (so a tray slides in and the coolant connections mate without hand plumbing), configurable I/O bays and fanless liquid-cooled compute trays.

250kW per rack is several times what air-cooled enterprise racks handle, which is why the design is liquid-first. The consequence for operators is that retrofitting older halls is hard; new builds and purpose-made AI halls benefit most. Our explainer on the real environmental impact of data centers covers why cooling choices matter for water and power.

NVIDIA also released a Vera CPU rack specification extending MGX to CPU-focused racks. The use case is agentic AI: tool calls, orchestration and sandboxed execution consume CPU cycles that sit beside the GPU racks. We analyzed the chip itself in our look at the Vera CPU whitepaper and its marketing claims.

DSX: open software for standing up an AI factory

NVIDIA open-sourced four DSX OS components, developed, it says, from its experience running DGX Cloud:

• Config Manager automates switch configuration.
• Cluster Readiness Engine stress-tests GPU clusters before training starts.
• Rack Management Service abstracts rack hardware for firmware management.
• Cluster API Provider for NVIDIA Infrastructure Controller speeds bare-metal provisioning; NVIDIA claims it turns hardware into production clusters in hours instead of days.

That claim is NVIDIA's own and unverified by third parties. Still, the pieces address a real pain: a cluster that arrives physically complete but takes weeks to qualify wastes expensive GPU time.

Networking and optics: Spectrum-6, Multiplane, USFP

Spectrum-6 switches move 102.4Tb/s with 200G per lane, twice the capacity and per-lane bandwidth of Spectrum-4. Spectrum-X Multiplane spreads GPU traffic across up to eight independent network planes, which NVIDIA says enables scale beyond 100,000 GPUs with fewer switch tiers. Meta is adopting Spectrum-6; its N-IceCube switch runs Meta's FBOSS network OS in the 102.4T IcePack liquid-cooled rack. Meta's Omar Baldonado, senior director of data center and AI networking, is quoted saying the company is excited about its partnership with NVIDIA "in making multiplane fabrics operationally scalable and performant."

On optics, USFP was co-announced with Cisco, with TE Connectivity and Amphenol working on cages, connectors and modules. It delivers 400Gb/s per lane, 16 full-duplex lanes scaling to 32, roughly four times the bandwidth density of 200Gb/s-per-lane pluggables and about 1.3x better power efficiency, in a footprint close to OSFP. NVIDIA also said Google's Falcon Transport will be supported on future ConnectX-10 SuperNICs and BlueField DPUs, adding multipath load balancing, programmable congestion control, hardware loss recovery and per-connection security.

Power and the grid: 800 VDC and the two-second rule

Two facility-level documents may matter more over time than any rack drawing.

• Open Data Center Specification v0.7, a facility blueprint for power, cooling and physical layout, co-published by NVIDIA with AMD, Google and Meta.
• OCP Grid Resilience Specification, published by Google, Meta, Microsoft and NVIDIA. It defines how AI facilities stay connected during grid disturbances and recover to near-full load in under two seconds.

Both are described as compatible with NVIDIA's DSX reference design. Partner Visionbay also unveiled K-1, an 800 VDC AI factory platform for future Vera Rubin deployments built on NVIDIA's 800 VDC reference architecture and Vertiv power and thermal technology, with cluster-level power centers and the option to phase 800 VDC in gradually.

Power cable linking a transformer to a server cabinet, representing grid resilience for Vera Rubin AI factories

Why the grid spec matters: AI training loads swing sharply, and utilities and regulators are pushing back on new campuses. Our AI data center backlash map and the hyperscaler nuclear deals show how power has become the binding constraint. A shared standard for riding through disturbances gives operators something concrete to offer a utility.

Where Vera Rubin stands today

This announcement is about designs and standards, not shipments. For deployment signals, see CoreWeave's first Vera Rubin cluster, the Nscale 100,000-GPU Vera Rubin deal with Figure and the SSI partnership. Component supply is a separate story: our report on GB300 packaging at TSMC Arizona and Amkor shows how packaging capacity gates output.

Why contribute designs to OCP at all?

The Open Compute Project began as a way for large operators to share hardware designs so that suppliers could build to a common standard instead of a custom one for each buyer. For a vendor, contributing a reference design does three things. It shortens the time from chip launch to a rack that a server maker can sell, because the mechanical, thermal and power envelope is already defined. It gives hyperscalers a document they can point to when negotiating with suppliers. And it makes the vendor's rack form factor the default that adjacent products, such as cooling distribution units, busbars and optics, are designed around.

MGX has been NVIDIA's route to that outcome: a modular reference that lets partners mix compute, networking and storage trays. MGX 1.2 pushes the same idea to liquid-cooled, very high power racks. The practical effect is that the choice facing a data center builder shifts from "which proprietary rack" to "which MGX-compatible supplier," and from a single integrator to a wider field of manufacturers.

There is also a timing logic. Facility lead times for power and cooling are long, often longer than chip lead times. Publishing the rack envelope early, along with the Open Data Center Specification and grid standards, lets operators start civil and electrical work for a platform that has not yet reached volume. That is useful to customers and, not coincidentally, helps lock in demand for the platform those facilities are built around.

Reading this critically

• Open is not free. Contributing to OCP lets others build compatible racks, but NVIDIA still controls the GPUs, NVLink and the software stack. Openness here lowers integration cost; it does not make the platform vendor-neutral.
• Spec versus silicon. A specification at v0.7 is a draft. Check which parts are final before committing capex.
• Performance numbers are NVIDIA's. Figures like 1.3x power efficiency for USFP and hours-not-days provisioning are vendor claims.
• Power is the hard part. 250kW racks and 800 VDC need new electrical gear, and local grid constraints can matter more than any rack standard.

What this means for builders and buyers

1. If you rent GPUs, expect Vera Rubin capacity to scale through more server makers and neocloud operators as designs spread, which should ease supply over time.
2. If you run a data center, plan for liquid cooling and higher power density from the start, and read the grid specification before talking to your utility.
3. If you build agents, note that NVIDIA is explicitly speccing CPU racks for orchestration and sandboxes; agent workloads are changing hardware ratios.
4. If you follow the supply chain, optics and connectors (Cisco, TE Connectivity, Amphenol) are now named parts of the plan.

Related reading

• CoreWeave's first Vera Rubin cluster
• Figure and Nscale 100,000-GPU Vera Rubin deal
• NVIDIA Vera CPU whitepaper critique
• AI data center backlash map
• Hyperscaler nuclear deals for AI data centers
• Real environmental impact of data centers
• Primary source: NVIDIA OCP Global Summit post; standards at the Open Compute Project

Specifications and partner announcements are accurate as of October 12, 2026 and come from NVIDIA's post.